Recognizing and Predicting Context by Learning from User Behavior

Transcription

1 Recognizing and Predicting Context by Learning from User Behavior 15. September 2003, Jakarta Institut für Praktische Informatik Johannes Kepler Universität Linz, Austria < 1 >

2 Vision A Personal Digital Assistant that can live up to its name. < 2 >

3 Mission Problem: Most information appliances are difficult to use for non-technology-savy users Devices only react to user input Users have to adapt to the devices Aim: Make information appliances smarter in a sense that they are easier to use Devices should be proactive Devices should adapt to the user < 3 >

4 Approach Personal information appliances should learn from user s habits Exploiting multiple sensors for context awareness Predicting future user context by learning from the past information appliance sensors (low level data) current user context user habits high level context information application future user context < 4 >

5 Content Introduction Context awareness Proactivity Architecture Step 0: Sensor data acquisition Step 1: Feature extraction Step 2: Classification Step 3: Labeling Step 4: Prediction First results < 5 >

6 Context awareness Many definitions for context, e.g. by Dey as any information that can be used to characterize the situation of an entity, where an entity can be a person, place or a physical or computational object Context has many aspects Using multiple simple sensors seems more reasonable to capture different aspects of context < 6 >

7 Proactivity < 7 >

8 Content Introduction Architecture Step 0: Sensor data acquisition Step 1: Feature extraction Step 2: Classification Step 3: Labeling Step 4: Prediction First results < 8 >

9 Architecture S 1 Feature extraction Classification Labeling Meeting n t Sensors S 2 S 3 C 1 C 2 At desk m t+ C m C k / P k S l Prediction <s 1, s 2,..., s l > t <f 1, f 2,..., f n > t <c 1, c 2,..., c m > t <c 1, c 2,..., c m > t+ Input vector (Sensor vector) Feature vector Class vector Future class vector < 9 >

10 Content Introduction Architecture Step 0: Sensor data acquisition Step 1: Feature extraction Step 2: Classification Step 3: Labeling Step 4: Prediction First results < 10 >

11 Sensors for (mobile) information appliances Typical sensors available for monitoring the user context: Time Application/Window manager Brightness Microphone Bluetooth Wireless LAN Docked / undocked Other suitable sensors can be connected: GPS GSM Compass Accelerometer Tilt sensor Temperature sensor Pressure sensor Sharing of sensor data between appliances < 11 >

12 Content Introduction Architecture Step 0: Sensor data acquisition Step 1: Feature extraction Step 2: Classification Step 3: Labeling Step 4: Prediction < 12 >

13 Feature Extraction Raw sensor data is transformed into more meaningful features Feature extraction exploits domain-specific knowledge Multiple features extracted from a single sensor High-dimensional input vectors Different types of features: Numerical (continuous) sensors: e.g. brightness sensor Numerical (discrete) sensors: e.g. number of access points in range Ordinal sensors: e.g. day of week Nominal sensors: e.g. WLAN-SSID, list of Bluetooth devices in spatial proximity Only two operations necessary for each feature: Distance metric Adaptation operator < 13 >

14 Feature extraction on a current PDA With current computational power easily possible Every feature implements distance metric and adaptation operator On-line classifications possible with any combination of features (user-selectable and loaded dynamically during startup) < 15 >

15 Content Introduction Architecture Step 0: Sensor data acquisition Step 1: Feature extraction Step 2: Classification Step 3: Labeling Step 4: Prediction First results < 16 >

16 Classification: Introduction Classifies feature vectors and finds common patterns in sensor data Different types of classification algorithms Type (partitioning / hierarchical) Soft / hard classification Supervised / unsupervised Requirements for classifying user context in information appliances: On-line learning Adaptivity Variable number of classes and variable topology Soft classification Noise resistance Limited resources Simplicity Interpretability of classes / protection of data privacy < 17 >

17 Classification: Algorithms Algorithm Network topology Topology preserving Competitive SOM fixed yes soft RSOM fixed yes soft K-Means fixed no hard Leader variable no hard Growing K-Means variable no hard Neural Gas variable no soft Neural Gas + Competitive Hebbian Learning variable yes soft Growing Neural Gas variable yes soft Incremental DBSCAN variable No hard < 18 >

18 Content Introduction Architecture Step 0: Sensor data acquisition Step 1: Feature extraction Step 2: Classification Step 3: Labeling Step 4: Prediction First results < 20 >

19 Assigning user-defined labels to user context t t t 1:{0,1} assignment of (meta-) clusters to context names Two possibilities: If the (meta-) clusters at the output of the classification step are stable, direct assignment to names If the (meta-) clusters are unstable due to learning and adaptation, use a second, simple classification step [van Laerhoven, 2001] ID 1 = at work ID 2 = meeting ID 3 = lunch ID 4 = ID 5 = at home < 21 >

20 Content Introduction Architecture Step 0: Sensor data acquisition Step 1: Feature extraction Step 2: Classification Step 3: Labeling Step 4: Prediction First results < 22 >

21 Prediction of User Context Recognized context classes can be regarded as states of an abstract state machine Monitoring the state trajectory allows to predict future states Prediction algorithm requirements for predicting context states : Unsupervised model estimation On-line learning Incremental model growing Confidence estimation Automatic feedback Manual feedback Long-term vs. short-term Architecture allows prediction algorithms to be realized as plug-ins Algorithms can be changed according to the specific application needs < 23 >

22 Content Introduction Architecture Step 0: Sensor data acquisition Step 1: Feature extraction Step 2: Classification Step 3: Labeling Step 4: Prediction First results < 24 >

23 First Results Test data gathered over 10 days for a selection of features: Timestamp Bluetooth: list of MAC addresses in spatial proximity Bluetooth: number of peers in spatial proximity Wireless LAN: list of MAC addresses in spatial proximity (clients associated to a nearby access point) Currently active application Distance metrics and adaptation operators implemented in straight-forward ways (e.g. Hamming distance for lists of MAC addresses) Implementation of Lifelong Growing Neural Gas (LLGNG) as classification algorithm < 25 >

24 First Results One-pass (online) run through the log file yields (multiple test runs): ~ 40 clusters ~ 6 meta-clusters < 0.12 overall classification error (average distance of input data points to cluster centers) < 26 >

25 Summary Architecture allows recognition and prediction of user context from raw sensor data in an un-supervised, non-obtrusive way in 4 steps: Feature extraction Classification Labeling Prediction Simple interfaces between the architecture layers Multiple, simple sensors instead of a single, complex one Heterogeneous feature vector handled via specific implementation of: distance metric adaptation operator for each feature First implementations of feature extraction and classification show promising results < 27 >

26 Conclusions and Future Outlook Proactive applications could allow the development of more intuitive user interfaces A Personal Digital Assistant and information appliances in general should become proactive to achieve a wider acceptance Context awareness is one possibility to achieve proactivity in applications Learning user behavior allows to predict future user context and thus future user actions Future research on non-obtrusive user interfaces for labeling user context necessary Qualitative comparison of prediction methods will be performed using realworld data < 28 >

27 It is hard to predict, especially the future Neils Bohr Winner of a Nobel Prize for physics < 29 >

28 The best way to predict the future is to invent it. Alan Kay Inventor of Smalltalk < 30 >

29 Thank you for your attention! < 31 >